APBRmetrics

A lot of people have pet theories about how different college coaches/systems impact the statistical output of their players thus biasing simple statistical projections. I assume there is a lot of value in this approach, but it is difficult to integrate into my projection models without serious concerns of overfitting. Still it would be nice to find some patterns consistent with this idea in order to help better interpret the simple model outputs.

I fit a draft projection model along with by-coach random slopes and random intercepts in order to look into these ideas and wanted to share the data here. The intercept column show how much players coming out of that system tend to under/over-perform relative to other coaches. Each of the other columns gives the by-coach random slope for each statistic, telling us how much less/more informative each box-score statistic is when collected in a particular coach's system relative to others.

Here are all coaches in the data to produce at least 10 NBA players ordered by intercept:



I you want to pull the full data and play with it, I posted it to my Google-doc here: http://bit.ly/1jomLsR

I'm really interested if anyone can find patterns like "coaches running a 2-3 zone have relatively low steal slopes" that may help us understand how different systems impact our stats-based assessment of players.

This was impressive and worthy of a bump re-view. I'll probably look more at it again around draft time.

This is awesome. So these intercepts are relative to your EWP target variable, or something else?

Not quite the same model as the EWP, but that is probably the best way to think of it.

I actually have something I did recently that gets at the same idea as this, but is a bit easier to work with. Link to the tool is here: https://laynevashro.shinyapps.io/coach_cors/

Rather than looking at the random effects, this is just using the correlation of each stats between leagues (adjusting for SOS and Age). It then allows you to highlight different coaches and position ranges to see how they vary from the baseline expectation.

Seems like a form of cross validation with a Bayesian prior of "no coach effect" should be able to give you some not-over-fitted results.

Yes. I like that idea. I've been thinking of some ways to pull this information into a useful framework. I don't know if any of you have seen my coaching network tool, but I made coaching networks based on assistantships that I am hoping to use to aggregate systems. Ideally, I would like to identify several core styles on O and D that could be statistically classifiable. I think I might fail at this, but if successful I could peg systems with priors based on coaching links, then aggregate across different systems to get useful model adjustments.

Thanks for revisiting this, it prompted me to start a series I've been meaning to work on for awhile. First post looking at Calipari is up at NylonCalculus. (link: http://nyloncalculus.com/2015/01/09/dra ... -calipari/)

It might be possible to over do coach influence analysis but it sounds good that you are going to explore the territory. First article is measured and helpful. If you like writing and showcasing your tools, this should be a deep vein to mine.